Event-Log Architecture Could Fix the Core Flaws in AI Agent Design

Most AI agent codebases, including Claude Code and Codex, share a common while-loop structure where a single mutable state object drives the entire session. This design creates recurring problems such as broken interruption handling, messy retry logic, awkward parallel tool calls, and no ability to rewind or replay sessions. A developer building a personal agent called Pizza proposed an alternative approach where an append-only event log serves as the sole source of truth, with state derived as a projection of that log. Every message, tool call, and file edit is stored as a row in an EventStore database, allowing the runtime to reconstruct context by reading the log rather than holding state in memory. This structure makes previously complex features like conversation branching, session replay, and long-running multi-day tasks straightforward by design rather than special-cased workarounds.